Learning a Control for Quadcopter with Reinforcement Learning

Development environment

WARNING The Dockerfile on this project uses a base image which has been hijacked.
The Dockerfile has been modified to comment out the hijacked base. If you wish to use the base image, please do so with caution.

A docker image is provided for the code development: brzl/quadcopter:devel. The definition of the image can be found inside the folder docker. As you can see, it is based on the brezel research platform, with additional python dependencies. The first thing you need to do is build the image, which can takes a while (but you usually need to build it once for all):

make build

The commands described in the rest of this README are supposed to be run from the devel container. Enter make run to get a shell.

If you need X11 suport, use make run-x11 instead. Do xrdb -merge Xresources once in the container.

Getting started

1. Build applications

bazel build :all

Note that you will need to run bazel sync --configure in case you modify the content of folder gym-quadcopter.

2. Run the training

bazel run :training

It will read params from config_training.yaml file. Results will be available in results/experiment_DATETIME.

3. Inspect results with Tensorboard

You need to enter the Tensorboard Container (from your host) with

make tensorboard

and then run Tensorboard making it point to the experiment directory

tensorboard --logdir /results/experiment_DATETIME/runs --bind_all

4. Export Weights for Sherlock

bazel run :export

It will read params from config_export.yaml file.

The input is the model/quadcopter-{i}.pkl selected and the export is defined in the config file.

Notes

• config_training.yaml has all hyperparameters
• in models: the pkl checkpoints
• in figures: all plots (rewards etc.)
• in log: log file
• in tests: all tests that are run with proper tags
• config_tests contains the corresponding parameters and figures and log inside each of them